The dense MoE kernels (moe_replicate, moe_bias_add_3d, moe_weighted_sum)
computed total / expert_stride / element indices in int32. gpt-oss prefill
runs the whole prompt through the dense path unchunked (SPARSE_MAX_TOKENS=8),
so local_experts*num_tokens*hidden (and batch*num_tokens*dim, and
local_id*expert_stride) overflow int32 at ~3.6k-23k prefill tokens
(TP-dependent) — well inside the supported context window. The launch then
fails silently because CUDA_CHECK_LAST_ERROR was ((void)0) under NDEBUG, so
the bias / weighted-sum simply never runs and the forward pass is corrupted
with no error reported.
Fix: switch the three kernels and their launchers to long long, mirroring the
(long long) indexing already used in moe_sparse.cu. Also make
CUDA_CHECK_LAST_ERROR always-on — cudaGetLastError does not sync, so the
per-launch host cost is negligible, and a silent launch failure is exactly
the class of bug this one was.
Verified on dash5 (RTX 5090): a direct kernel test at 2.21B elements (>2^31)
for both moe_replicate and moe_bias_add_3d produces correct results with no
launch error; bench-gpt-oss TP=2 holds at 5.9ms TPOT, output unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- --pp with gpt-oss now fails with a clear message instead of a
cryptic missing-weight panic inside the Qwen3-only PP engine.
- Sparse GEMV wrappers assert K%16==0 (FP8) / K%32==0 (MXFP4) — the
uint4-vectorized kernels would silently drop a tail otherwise.
- Document the topk_ids buffer holding i32 under an F32 dtype label
(DType has no I32).
- Drop unused imports/locals and the cuBLASLt scale-mode constants
orphaned by the strided-batched FP8 rework (e631a71).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dense MoE replicated x across all 16 local experts and ran the full
batched GEMM, reading every expert's weights per token; the weighted
sum then discarded 12 of 16 results. Decode is memory-bound, so this
was ~8x wasted expert bytes — the entire decode gap vs llama.cpp.
New fused expert-indexed GEMVs (csrc/moe/moe_sparse.cu) read
topk_ids on-device (no host sync) and early-return block-uniformly
for experts other ranks own. FP8 runs W8A16 (activations stay BF16 —
tensor cores are irrelevant at M=1, and activation quantization error
disappears); MXFP4 runs W4A16. Per-expert bias + scale fused into the
GEMV epilogue; slot-indexed weighted sum skips (never multiplies)
unwritten non-local slots. Dense path retained for num_tokens > 8
(prefill) and via XSERV_DENSE_MOE=1 for A/B.
dash5 (RTX 5090), gpt-oss-20b FP8, TP=2: decode TPOT 13.9 -> 7.6 ms.
Warm-server vs llama.cpp MXFP4 TP=2: TPOT 7.19-7.32 vs 7.54-8.42 ms —
first config where xserv wins decode outright. GSM8K-100: 96% (dense
FP8: 91%). llama TP=1 (2.9 ms) remains ahead: next levers are decode
CUDA graphs, non-expert quantization, sparse prefill (docs/20).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replace the per-token CPU-routed MoE forward with an all-GPU path:
1. moe_topk_softmax: GPU top-k + softmax (was CPU sort + softmax)
2. moe_replicate: broadcast input to all local experts
3. cublasGemmStridedBatchedEx: batched expert matmul (was per-expert cuBLAS)
4. moe_weighted_sum: FP32-accumulated weighted sum on GPU (was GPU→CPU→F32→BF16→GPU)
Expert weights stored as contiguous 3D tensors for strided batched GEMM.
Zero CPU↔GPU transfers per MoE layer (was ~40 per token per layer).
Also: configurable geglu_alpha, LayerNorm bias auto-detect, unused-weight
diagnostic at load time.
GSM8K 30-problem: 11/30 → 23/30 (76.7%) vs llama.cpp 30/30 (100%).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
